Notebook 1 — From raw Geolife trajectories to visit-level events

This notebook shows how raw Geolife GPS trajectories for one user are turned into a cleaned visit-level dataset. I load and filter the original .plt files, restrict the data to Beijing, and project all points onto a 500 m grid. Night-time points are used to infer the user’s home (HOME) and a secondary frequently visited place (SW), and a simple movement-based rule selects additional grid cells as daily activity locations. Consecutive points within the same activity cell are then collapsed into single visit events and labelled as HOME, SW, first-time place (Pv), or return place (Pn). The resulting visit-level table is the starting point for all subsequent analyses.

1. Environment and data paths

The Geolife data folder Geolife Trajectories 1.3 is stored in the same directory as this notebook. Below I set the base directory and import the Python packages used throughout the analysis.

import os
import glob
import math

import numpy as np
import pandas as pd
import pyproj

# Path to the Geolife data (relative to the notebook)
BASE_DIR = os.path.join("Geolife Trajectories 1.3", "Data")

# Example user (we start with user 000)
USER_ID = "000"

# Optional: restrict to Beijing area
FILTER_BEIJING = True
BEIJING_BBOX = (115.42, 39.44, 117.50, 41.06)   # (minLon, minLat, maxLon, maxLat)

# Column names for Geolife .plt files (after skipping the 6-line header)
PLT_COLS = ["lat", "lon", "unused", "altitude_feet", "days", "date", "time"]

2. Inspecting raw GPS trajectories

Before any spatial filtering or grid construction, the raw Geolife records for user 000 are inspected. Table 1 previews the cleaned point-level data: each row corresponds to a single GPS fix with a timestamp, derived calendar date and hour, latitude/longitude, and altitude.

The spatial footprint of these records is visualised by plotting the raw trajectories as daily polylines in geographic coordinates (Fig. 2). Connecting consecutive points into lines highlights dense clusters of movement as well as the long-distance corridor between Beijing and Shanghai, and motivates the later projection to UTM and discretisation onto a 500 m grid.

Temporal coverage is assessed by aggregating the data by date and counting the number of recorded GPS points per day (Fig. 3). The resulting time series clearly shows highly irregular tracking: some days contain several thousand points, others only a few, and there are extended gaps with no data at all. This pattern reflects the device being switched on only intermittently and at changing sampling rates, a key limitation that subsequent behavioural analyses must be interpreted conditional on.

def read_one_plt(path, user_id="000"):
    """Read a single Geolife .plt file, clean basic issues, and return a DataFrame."""
    df = pd.read_csv(path, skiprows=6, names=PLT_COLS)

    # Basic cleaning: drop rows with missing coords or time
    df = df[
        pd.notnull(df["lat"]) &
        pd.notnull(df["lon"]) &
        pd.notnull(df["date"]) &
        pd.notnull(df["time"])
    ].copy()

    # Build timestamp
    df["datetime"] = pd.to_datetime(
        df["date"].astype(str) + " " + df["time"].astype(str),
        errors="coerce"
    )
    df = df[pd.notnull(df["datetime"])].copy()

    # Derive date and hour
    df["user"] = user_id
    df["file"] = os.path.basename(path)
    df["date_only"] = df["datetime"].dt.date
    df["hour"] = df["datetime"].dt.hour

    # Keep core columns
    df = df[["user", "file", "datetime", "date_only", "hour",
             "lat", "lon", "altitude_feet"]]
    return df


# Scan all trajectory files for the chosen user
traj_glob = os.path.join(BASE_DIR, USER_ID, "Trajectory", "*.plt")
files = sorted(glob.glob(traj_glob))

dfs = []
for fp in files:
    try:
        dfs.append(read_one_plt(fp, user_id=USER_ID))
    except Exception as e:
        print(f"[warning] failed to read {fp}: {e}")

if not dfs:
    raise RuntimeError("No trajectories were successfully read. Check the data path.")

traj = pd.concat(dfs, ignore_index=True)
traj.sort_values("datetime", inplace=True)
traj.reset_index(drop=True, inplace=True)
traj.head(10)

	user	file	datetime	date_only	hour	lat	lon	altitude_feet
0	000	20081023025304.plt	2008-10-23 02:53:04	2008-10-23	2	39.984702	116.318417	492
1	000	20081023025304.plt	2008-10-23 02:53:10	2008-10-23	2	39.984683	116.318450	492
2	000	20081023025304.plt	2008-10-23 02:53:15	2008-10-23	2	39.984686	116.318417	492
3	000	20081023025304.plt	2008-10-23 02:53:20	2008-10-23	2	39.984688	116.318385	492
4	000	20081023025304.plt	2008-10-23 02:53:25	2008-10-23	2	39.984655	116.318263	492
5	000	20081023025304.plt	2008-10-23 02:53:30	2008-10-23	2	39.984611	116.318026	493
6	000	20081023025304.plt	2008-10-23 02:53:35	2008-10-23	2	39.984608	116.317761	493
7	000	20081023025304.plt	2008-10-23 02:53:40	2008-10-23	2	39.984563	116.317517	496
8	000	20081023025304.plt	2008-10-23 02:53:45	2008-10-23	2	39.984539	116.317294	500
9	000	20081023025304.plt	2008-10-23 02:53:50	2008-10-23	2	39.984606	116.317065	505

import folium
traj_for_map = traj.copy()

# Center the map at the median lat/lon of all points
center_lat = traj_for_map["lat"].median()
center_lon = traj_for_map["lon"].median()

m_raw = folium.Map(
    location=[center_lat, center_lon],
    zoom_start=8,
    tiles="cartodbpositron"
)

# Draw one polyline per day to show daily trajectories
for d, sub in traj_for_map.groupby("date_only"):
    sub = sub.sort_values("datetime")
    if len(sub) < 2:
        continue

    coords = list(zip(sub["lat"].values, sub["lon"].values))  # (lat, lon) pairs
    folium.PolyLine(
        locations=coords,
        weight=1,
        opacity=0.4
    ).add_to(m_raw)

m_raw

Make this Notebook Trusted to load map: File -> Trust Notebook

# Daily number of GPS points
day_counts = (
    traj
    .groupby("date_only")
    .size()
    .rename("n_points")
    .reset_index()
)

# Simple time-series style plot of counts per day
day_counts_plot = day_counts.copy()
day_counts_plot["date_only"] = pd.to_datetime(day_counts_plot["date_only"])

import matplotlib.pyplot as plt

plt.figure(figsize=(7, 3.5))
plt.plot(
    day_counts_plot["date_only"],
    day_counts_plot["n_points"],
    marker="o",
    linewidth=1
)
plt.xlabel("Date")
plt.ylabel("Number of GPS points")
plt.title("Daily number of recorded GPS points for user 000")
plt.tight_layout()
plt.show()

3. Spatial filter: restricting to the Beijing area

Geolife users may have trajectories in multiple cities. In this project we focus on the Beijing region. We therefore apply a simple bounding box filter on longitude and latitude.

if FILTER_BEIJING:
    minLon, minLat, maxLon, maxLat = BEIJING_BBOX
    before = len(traj)
    traj = traj[
        (traj["lon"] >= minLon) & (traj["lon"] <= maxLon) &
        (traj["lat"] >= minLat) & (traj["lat"] <= maxLat)
    ].copy()
    after = len(traj)
    print(f"Beijing filter: {before} -> {after} points")
else:
    print("Beijing filter is disabled.")

Beijing filter: 173870 -> 157646 points

4. Projection to UTM and construction of a 500 m grid

All GPS points are transformed from WGS84 geographic coordinates (lat/lon) to UTM Zone 50N so that distances can be measured in metres. In this projected space, a regular 500 m × 500 m grid is built to fully cover the user’s trajectory. Each GPS fix is then mapped to its corresponding grid cell (ci, rj).

This discretisation step reduces the influence of GPS jitter and avoids tracing every minor wiggle of the raw trajectories. Subsequent analyses operate at the cell level—using visits to grid cells instead of raw points—which stabilises behaviour patterns and prevents over-interpreting micro-movements that arise from measurement error or inconsistent sampling.

Using a fixed grid inevitably introduces a mild modifiable areal unit problem (MAUP): the exact boundaries and grid size may slightly influence which cells a point falls into. In this context, however, the 500 m resolution strikes a balance between reducing noise and retaining meaningful activity locations, making the grid a practical spatial unit for behavioural modelling.

def cell_center_lonlat(ci, rj):
    cx = gx0 + (ci + 0.5) * GRID_SIZE
    cy = gy0 + (rj + 0.5) * GRID_SIZE
    lon, lat = to_wgs(cx, cy)
    return lon, lat

# Coordinate reference systems
CRS_WGS84 = pyproj.CRS("EPSG:4326")
CRS_UTM50 = pyproj.CRS("EPSG:32650")
to_utm = pyproj.Transformer.from_crs(CRS_WGS84, CRS_UTM50, always_xy=True).transform
to_wgs = pyproj.Transformer.from_crs(CRS_UTM50, CRS_WGS84, always_xy=True).transform

# Lat/lon -> projected (metres)
x, y = to_utm(traj["lon"].values, traj["lat"].values)
traj["x"] = x
traj["y"] = y

GRID_SIZE = 500  # metres

# Grid extent with one extra cell of padding on each side
minx, miny = traj["x"].min(), traj["y"].min()
maxx, maxy = traj["x"].max(), traj["y"].max()

gx0 = math.floor(minx / GRID_SIZE) * GRID_SIZE - GRID_SIZE
gy0 = math.floor(miny / GRID_SIZE) * GRID_SIZE - GRID_SIZE
gx1 = math.ceil (maxx / GRID_SIZE) * GRID_SIZE + GRID_SIZE
gy1 = math.ceil (maxy / GRID_SIZE) * GRID_SIZE + GRID_SIZE

ncol = int((gx1 - gx0) / GRID_SIZE)
nrow = int((gy1 - gy0) / GRID_SIZE)
print(f"Grid columns × rows: {ncol} × {nrow} (total {ncol*nrow:,} cells)")

# Assign each point to a grid cell
traj["ci"] = ((traj["x"] - gx0) // GRID_SIZE).astype(int)
traj["rj"] = ((traj["y"] - gy0) // GRID_SIZE).astype(int)

traj[["datetime", "lat", "lon", "x", "y", "ci", "rj"]].head()

Grid columns × rows: 128 × 102 (total 13,056 cells)

	datetime	lat	lon	x	y	ci	rj
0	2008-10-23 02:53:04	39.984702	116.318417	441807.056623	4.426282e+06	43	50
1	2008-10-23 02:53:10	39.984683	116.318450	441809.858037	4.426280e+06	43	50
2	2008-10-23 02:53:15	39.984686	116.318417	441807.043048	4.426280e+06	43	50
3	2008-10-23 02:53:20	39.984688	116.318385	441804.312590	4.426280e+06	43	50
4	2008-10-23 02:53:25	39.984655	116.318263	441793.868245	4.426277e+06	43	50

5. Detecting HOME and SW from night-time points

I infer HOME and a secondary frequent place (SW) from points observed between 00:00–06:00 on a 500 m grid, using night-time frequency as a proxy for home-like locations. HOME is the cell with the most night points (ties broken by longer time span), and SW is the next strongest night-time cell after removing HOME, if available.

night = traj[(traj["hour"] >= 0) & (traj["hour"] < 6)].copy()
night["cell"] = list(zip(night["ci"], night["rj"]))
print("Number of night-time points:", len(night))

cell_counts = night["cell"].value_counts()

Number of night-time points: 47296

def night_span_seconds(cell):
    sub = night[night["cell"] == cell]["datetime"]
    return (sub.max() - sub.min()).total_seconds() if not sub.empty else 0

# HOME: cell with the largest night-time count; if tied, pick the one with largest time span
candidates = cell_counts[cell_counts == cell_counts.max()].index.tolist()
home_cell = max(candidates, key=night_span_seconds) if len(candidates) > 1 else cell_counts.index[0]

# SW: second-strongest night-time cell after removing HOME
cell_counts_wo_home = cell_counts[cell_counts.index != home_cell]
sw_cell = None
if not cell_counts_wo_home.empty:
    cand2 = cell_counts_wo_home[cell_counts_wo_home == cell_counts_wo_home.max()].index.tolist()
    sw_cell = max(cand2, key=night_span_seconds) if len(cand2) > 1 else cell_counts_wo_home.index[0]

home_lon, home_lat = cell_center_lonlat(*home_cell)
print(f"HOME cell = {home_cell} @ ({home_lon:.6f}, {home_lat:.6f})")

if sw_cell is not None:
    sw_lon, sw_lat = cell_center_lonlat(*sw_cell)
    print(f"SW   cell = {sw_cell} @ ({sw_lon:.6f}, {sw_lat:.6f})")
else:
    print("SW   cell = None")

HOME cell = (44, 55) @ (116.323385, 40.006970)
SW   cell = (43, 55) @ (116.317527, 40.006935)

6. Daily activity cells via line–grid intersections

Not every grid cell that contains a raw GPS point should be treated as an “activity location.” Points recorded along fast road segments are usually in transit rather than true stops. To produce a more conservative set of daily activity cells, I work with line segments instead of points: for each day, I connect consecutive projected points into segments in the (x, y) plane and count, for each grid cell, how many segments intersect its polygon.

A cell is labeled an activity location if its intersection count exceeds a daily threshold. I start with a baseline of 100 crossings, which filters most purely in-transit cells while retaining dense clusters around home and other frequently visited areas. This value is chosen empirically by inspecting multiple days: lower thresholds misclassify long road stretches as “places,” while higher thresholds begin to drop plausible stops. The threshold is then adapted day by day to avoid unrealistically “busy” days: if more than 30 cells exceed the threshold on a given day, I increase the threshold and recompute the classification.

from shapely.geometry import LineString, box
from shapely.strtree import STRtree

# Parameters for the line–grid crossing rule
THRESH_START = 100          # initial threshold for "enough crossings"
THRESH_STEP  = 25          # how much to increase the threshold if a day has too many cells
THRESH_MAX   = 1000        # upper bound on the threshold

MAX_PLACES_PER_DAY = 30    # per day: at most this many DISTINCT activity cells (places)

# Cache for grid-cell polygons (speeds things up)
_poly_cache = {}

def cell_poly(ci, rj):
    """Return the shapely Polygon for grid cell (ci, rj)."""
    key = (ci, rj)
    if key not in _poly_cache:
        x0 = gx0 + ci * GRID_SIZE
        y0 = gy0 + rj * GRID_SIZE
        _poly_cache[key] = box(x0, y0, x0 + GRID_SIZE, y0 + GRID_SIZE)
    return _poly_cache[key]

def query_segments(tree, poly, segs):
    """
    Wrapper around STRtree.query to handle both 'index' and 'geometry' return types,
    depending on Shapely version.
    """
    res = tree.query(poly)
    if len(res) == 0:
        return []
    first = res[0]
    # Some Shapely versions return indices, some return geometries
    if isinstance(first, (int, np.integer)):
        return [segs[i] for i in res]
    return res

# Containers for results
visited_cells_by_day = {}   # date -> set of (ci, rj) cells considered "activity locations"
cross_threshold_used = {}   # date -> threshold value actually used for that day

# We iterate over each date present in the trajectory
dates = sorted(traj["date_only"].unique().tolist())
print(f"Number of days with data for user {USER_ID}: {len(dates)}")

for d in dates:
    # All points for this day, ordered in time
    day = traj[traj["date_only"] == d].sort_values("datetime").copy()
    if len(day) < 2:
        # Not enough points to form line segments
        visited_cells_by_day[d] = set()
        cross_threshold_used[d] = THRESH_START
        continue

    # Build line segments between consecutive points (in projected coordinates)
    pts = list(zip(day["x"].values, day["y"].values))
    segs = [LineString([pts[i], pts[i+1]]) for i in range(len(pts) - 1)]

    # Drop empty / zero-length segments
    segs = [s for s in segs if (not s.is_empty) and (s.length > 0)]
    if not segs:
        visited_cells_by_day[d] = set()
        cross_threshold_used[d] = THRESH_START
        continue

    tree = STRtree(segs)

    # Only check the grid cells that actually appear for this day
    cmin, cmax = int(day["ci"].min()), int(day["ci"].max())
    rmin, rmax = int(day["rj"].min()), int(day["rj"].max())

    # Start from the base threshold and adapt if needed
    thr = THRESH_START
    while True:
        today_cells = set()

        # Loop over all relevant grid cells for that day
        for ci in range(cmin, cmax + 1):
            for rj in range(rmin, rmax + 1):
                poly = cell_poly(ci, rj)
                cnt = 0

                # Candidate segments intersecting this cell
                for seg in query_segments(tree, poly, segs):
                    if seg.intersects(poly):
                        cnt += 1
                        if cnt >= thr:
                            today_cells.add((ci, rj))
                            break   # no need to count further for this cell

        # If the day has too many activity CELLS, raise the threshold and try again
        if len(today_cells) > MAX_PLACES_PER_DAY and thr < THRESH_MAX:
            thr = min(thr + THRESH_STEP, THRESH_MAX)
        else:
            visited_cells_by_day[d] = today_cells
            cross_threshold_used[d] = thr
            break

# Summarise: how many activity cells per day, and what threshold was used
perday_summary = (
    pd.DataFrame({
        "date": [str(d) for d in dates],
        "visited_cells": [len(visited_cells_by_day[d]) for d in dates],
        "threshold_used": [cross_threshold_used[d] for d in dates],
    })
    .sort_values("date")
    .reset_index(drop=True)
)

print("Daily activity-cell counts (first 10 days):")
perday_summary.head(10)

Number of days with data for user 000: 122
Daily activity-cell counts (first 10 days):

	date	visited_cells	threshold_used
0	2008-10-23	3	100
1	2008-10-24	1	100
2	2008-10-26	2	100
3	2008-10-27	0	100
4	2008-10-28	3	100
5	2008-10-29	0	100
6	2008-11-03	0	100
7	2008-11-04	4	100
8	2008-11-10	0	100
9	2008-11-11	4	100

7. From daily activity cells to visit-level events

Given the daily activity cells, I next separate first-time places from returns and compress trajectories into visit-level events. Over the full observation window, a grid cell is labeled a first-time place (Pv) on the first date it appears as an activity cell; on any later day when it is active again, it is treated as a return place (Pn). HOME and SW remain separate categories and override Pv/Pn labels.

Using these labels, I build the visit-level table day by day. For each calendar day, I first compress the sequence of visited grid cells using simple run-length encoding, so consecutive points in the same cell become a single block. Each block is matched to that day’s activity-cell set and tagged as HOME, SW, Pv, or Pn; blocks not classified as activity locations are dropped as in-transit segments. Adjacent blocks with the same cell and tag are then merged so brief GPS gaps do not split a single stay into multiple visits.

The remaining blocks define visit events. For each visit, I record its start/end time, duration, grid indices and cell center (in both projected and geographic coordinates), distance to HOME, and a place identifier (HOME, SW, Pv#, Pn#). I also keep two within-day order measures: a running visit order across the day, and an action-order index that restarts at 1 when the user leaves HOME and resets to 0 upon returning. Finally, each visit receives a next_step label indicating whether the next visit is to HOME, SW, a Pv, a Pn, or none (end of day). Stacking all days yields the final visit-level table, saved as visit_level_table_000.csv and used as input for the subsequent notebooks.

from datetime import timedelta

pv_label = {}              # (ci, rj) -> "Pv#"
pn_label = {}              # ((ci, rj), date) -> "Pn#"
first_visit_date = {}      # (ci, rj) -> first date when this cell is an activity cell

pv_counter = 0
pn_counter = 0

for d in sorted(visited_cells_by_day.keys()):
    cells = visited_cells_by_day[d]
    for cell in sorted(cells):
        if cell not in first_visit_date:
            # First time this activity cell appears in the whole sample -> Pv
            pv_counter += 1
            first_visit_date[cell] = d
            pv_label[cell] = f"Pv{pv_counter}"
        else:
            # Subsequent days when the same activity cell is active -> Pn
            pn_counter += 1
            pn_label[(cell, d)] = f"Pn{pn_counter}"

print(f"Total unique Pv cells: {len(pv_label)}")
print(f"Total Pn labels assigned: {len(pn_label)}")


# --- 7.2 Helper: classify a cell on a given day ---

def classify_cell(cell, day_date):
    """
    Classify a grid cell on a given date into:
      - ('home', 'HOME')
      - ('sw', 'SW')
      - ('pv', 'Pv#')
      - ('pn', 'Pn#')
      - ('none', '')  for cells that are not treated as activity locations on that day.
    """
    if cell == home_cell:
        return "home", "HOME"
    if (sw_cell is not None) and (cell == sw_cell):
        return "sw", "SW"

    visited_today = visited_cells_by_day.get(day_date, set())
    if cell in visited_today:
        if first_visit_date.get(cell) == day_date:
            return "pv", pv_label[cell]
        else:
            return "pn", pn_label.get((cell, day_date), "PN")

    return "none", ""


def cell_center_xy(ci, rj):
    """Return the projected (x, y) centre of grid cell (ci, rj)."""
    cx = gx0 + (ci + 0.5) * GRID_SIZE
    cy = gy0 + (rj + 0.5) * GRID_SIZE
    return cx, cy


# --- 7.3 Build the merged visit-level table ---

visit_rows = []
all_dates = sorted(traj["date_only"].unique().tolist())

for d in all_dates:
    day = traj[traj["date_only"] == d].sort_values("datetime").copy()
    if day.empty:
        continue

    # Sequence of grid cells and timestamps for this day
    cells = list(zip(day["ci"].astype(int), day["rj"].astype(int)))
    times = day["datetime"].tolist()

    # 1) Raw run-length encoding by grid cell:
    #    consecutive identical (ci, rj) are grouped into one block.
    runs = []  # list of (cell, i0, i1) with indices into 'times'
    if cells:
        start = 0
        for i in range(1, len(cells)):
            if cells[i] != cells[i-1]:
                runs.append((cells[i-1], start, i-1))
                start = i
        runs.append((cells[-1], start, len(cells) - 1))

    # 2) Attach labels and time bounds; drop blocks that are not activity cells
    tagged = []
    for (cell, i0, i1) in runs:
        tag, pid = classify_cell(cell, d)
        if tag == "none":
            continue  # ignore transit-only cells
        tagged.append({
            "cell": cell,
            "ci": cell[0],
            "rj": cell[1],
            "tag": tag,
            "place_id": pid,
            "i0": i0,
            "i1": i1,
            "first_dt": times[i0],
            "last_dt": times[i1],
        })

    if not tagged:
        continue

    # 3) Merge adjacent blocks with the same (ci, rj, tag)
    merged = []
    current = tagged[0]
    for nxt in tagged[1:]:
        if (
            (nxt["ci"] == current["ci"]) and
            (nxt["rj"] == current["rj"]) and
            (nxt["tag"] == current["tag"])
        ):
            # extend the current block
            current["i1"] = nxt["i1"]
            current["last_dt"] = nxt["last_dt"]
        else:
            merged.append(current)
            current = nxt
    merged.append(current)

    # 4) For each merged visit, compute attributes and within-day order
    visit_order = 0     
    action_order = 0    

    for k, rec in enumerate(merged):
        ci, rj = rec["ci"], rec["rj"]
        tag, pid = rec["tag"], rec["place_id"]
        first_dt, last_dt = rec["first_dt"], rec["last_dt"]

        visit_order += 1

        if tag == "home":
            action_order = 0    
            action_ord = 0
        else:
            if action_order == 0:
                action_order = 1  
            else:
                action_order += 1
            action_ord = action_order

        # Next-step label (based on the next merged block)
        if k < len(merged) - 1:
            next_tag = merged[k + 1]["tag"]
        else:
            next_tag = "none"

        # Geometry-based attributes
        cx, cy = cell_center_xy(ci, rj)
        hx, hy = cell_center_xy(*home_cell)
        dist_home = math.hypot(cx - hx, cy - hy)
        center_lon, center_lat = cell_center_lonlat(ci, rj)

        visit_rows.append({
            "person": USER_ID,
            "date": d.isoformat(),
            "first_time": first_dt.time().isoformat(),
            "last_time": last_dt.time().isoformat(),
            "duration_min": round((last_dt - first_dt) / timedelta(minutes=1), 1),
            "grid_ci": ci,
            "grid_rj": rj,
            "grid_center_lon": round(center_lon, 6),
            "grid_center_lat": round(center_lat, 6),
            "dist_home_m": round(dist_home, 1),
            "is_home": 1 if tag == "home" else 0,
            "is_sw":   1 if tag == "sw"   else 0,
            "is_pv":   1 if tag == "pv"   else 0,
            "is_pn":   1 if tag == "pn"   else 0,
            "place_id": pid,          # HOME / SW / Pv# / Pn#
            "next_step": next_tag,    # home / sw / pv / pn / none
            "visit_order_in_day": visit_order,  
            "action_order": action_ord         
        })

# Final visit-level table
visit_table = (
    pd.DataFrame(visit_rows)
    .sort_values(["date", "first_time", "grid_ci", "grid_rj"])
    .reset_index(drop=True)
)

print("Number of visit events:", len(visit_table))
visit_table.head(10)

Total unique Pv cells: 106
Total Pn labels assigned: 294
Number of visit events: 1841

	person	date	first_time	last_time	duration_min	grid_ci	grid_rj	grid_center_lon	grid_center_lat	dist_home_m	is_home	is_sw	is_pv	is_pn	place_id	next_step	visit_order_in_day	action_order
0	000	2008-10-23	03:00:55	04:13:32	72.6	40	50	116.300184	39.984308	3201.6	0	0	1	0	Pv1	sw	1	1
1	000	2008-10-23	09:42:25	09:42:30	0.1	43	55	116.317527	40.006935	500.0	0	1	0	0	SW	home	2	2
2	000	2008-10-23	09:42:35	10:05:29	22.9	44	55	116.323385	40.006970	0.0	1	0	0	0	HOME	sw	3	0
3	000	2008-10-23	10:05:34	10:30:15	24.7	43	55	116.317527	40.006935	500.0	0	1	0	0	SW	home	4	1
4	000	2008-10-23	10:30:20	11:10:57	40.6	44	55	116.323385	40.006970	0.0	1	0	0	0	HOME	none	5	0
5	000	2008-10-24	02:09:59	02:10:54	0.9	43	55	116.317527	40.006935	500.0	0	1	0	0	SW	home	1	1
6	000	2008-10-24	02:10:59	02:47:06	36.1	44	55	116.323385	40.006970	0.0	1	0	0	0	HOME	none	2	0
7	000	2008-10-26	14:04:27	14:12:42	8.2	55	33	116.388704	39.908225	12298.4	0	0	1	0	Pv4	pv	1	1
8	000	2008-10-26	14:23:42	14:35:17	11.6	56	31	116.394633	39.899246	13416.4	0	0	1	0	Pv5	none	2	2
9	000	2008-10-27	12:03:59	12:05:54	1.9	44	55	116.323385	40.006970	0.0	1	0	0	0	HOME	none	1	0

# save visit-level table for later notebooks
OUT_DIR = "data"
os.makedirs(OUT_DIR, exist_ok=True)

out_path = os.path.join(OUT_DIR, f"visit_level_table_{USER_ID}.csv")
visit_table.to_csv(out_path, index=False, encoding="utf-8-sig")

8. Mapping inferred activity places

To summarise the cleaned visit-level data, I project the inferred activity locations back onto an interactive map. Starting from the visit table, I collapse repeated visits to the same 500 m grid cell and draw one square for each distinct cell that ever appears as a visit. Each cell is coloured according to its role in the visit history (explored at least once versus only revisited), while the HOME and SW cells are highlighted separately as point markers. Compared with the raw trajectories, this map shows a much simpler picture of the user’s daily activity space: a small set of recurrent places organised around home, rather than every individual GPS point and transit segment.

# --- 8.1 Aggregate unique places from the visit-level table ---

# We keep only non-home visits (HOME is shown separately as a point),
# and collapse repeated visits to the same grid cell.
places = (
    visit_table
    .groupby(["grid_ci", "grid_rj"], as_index=False)
    .agg(
        any_home=("is_home", "max"),
        any_sw=("is_sw", "max"),
        any_pv=("is_pv", "max"),
        any_pn=("is_pn", "max"),
        grid_center_lon=("grid_center_lon", "first"),
        grid_center_lat=("grid_center_lat", "first"),
        dist_home_m=("dist_home_m", "min")
    )
)


print("Number of unique grid cells appearing in visits:", len(places))


# --- 8.2 Helper: cell polygon in lat/lon ---

def cell_bounds_lonlat(ci, rj):
    """
    Return the 4 corners (and closed ring) of grid cell (ci, rj)
    as (lat, lon) pairs for use in folium.Polygon.
    """
    x0 = gx0 + ci * GRID_SIZE
    y0 = gy0 + rj * GRID_SIZE
    x1 = x0 + GRID_SIZE
    y1 = y0 + GRID_SIZE

    lon0, lat0 = to_wgs(x0, y0)
    lon1, lat1 = to_wgs(x1, y0)
    lon2, lat2 = to_wgs(x1, y1)
    lon3, lat3 = to_wgs(x0, y1)

    # folium expects (lat, lon)
    return [
        (lat0, lon0),
        (lat1, lon1),
        (lat2, lon2),
        (lat3, lon3),
        (lat0, lon0),
    ]


# --- 8.3 Colour rule for places ---

def color_for_place(row):
    cell = (int(row["grid_ci"]), int(row["grid_rj"]))
    # HOME and SW will be shown as point markers; here we colour only squares
    if sw_cell is not None and cell == sw_cell:
        return "green"
    # Pv vs Pn logic
    if row["any_pn"] == 1:
        return "orange"   # ever seen as Pv at least once
    if row["any_pv"] == 1:
        return "purple"   # only returns
    return "gray"


# --- 8.4 Initialise a folium map centred at HOME ---

# Compute HOME centre in lat/lon (from grid indices)
hx = gx0 + (home_cell[0] + 0.5) * GRID_SIZE
hy = gy0 + (home_cell[1] + 0.5) * GRID_SIZE
home_lon, home_lat = to_wgs(hx, hy)

m = folium.Map(location=[home_lat, home_lon],
               zoom_start=12,
               tiles="cartodbpositron")


# --- 8.5 Add grid-cell polygons for activity places ---

for _, row in places.iterrows():
    ci = int(row["grid_ci"])
    rj = int(row["grid_rj"])
    poly_latlon = cell_bounds_lonlat(ci, rj)
    col = color_for_place(row)

    popup_html = (
        f"Cell: ({ci}, {rj})<br>"
        f"Center: ({row['grid_center_lat']:.5f}, {row['grid_center_lon']:.5f})<br>"
        f"Dist. to HOME: {int(round(row['dist_home_m']))} m<br>"
        f"Any Pv: {int(row['any_pv'])} &nbsp; Any Pn: {int(row['any_pn'])}"
    )

    folium.Polygon(
        locations=poly_latlon,
        color=col,
        weight=2,
        fill=True,
        fill_opacity=0.35,
        popup=popup_html
    ).add_to(m)


# --- 8.6 Add HOME and SW as point markers ---

folium.CircleMarker(
    location=[home_lat, home_lon],
    radius=7,
    color="blue",
    fill=True,
    fill_opacity=0.9,
    popup="HOME"
).add_to(m)

if sw_cell is not None:
    sx = gx0 + (sw_cell[0] + 0.5) * GRID_SIZE
    sy = gy0 + (sw_cell[1] + 0.5) * GRID_SIZE
    sw_lon, sw_lat = to_wgs(sx, sy)

    folium.CircleMarker(
        location=[sw_lat, sw_lon],
        radius=7,
        color="green",
        fill=True,
        fill_opacity=0.9,
        popup="SW"
    ).add_to(m)


m

Number of unique grid cells appearing in visits: 106

Make this Notebook Trusted to load map: File -> Trust Notebook